This website uses cookies. By using the site you are agreeing to our Privacy Policy.



Fujifilm developed the world's first technology capable of searching and designing new drug candidate compounds based only on the structural formula of a known biologically active compound

Newly developed AI and simulation technology to automatically design new drug candidate compounds significantly increasing drug development success rates

October 4, 2018

FUJIFILM Corporation (President: Kenji Sukeno) has developed the world's first AI and simulation technology (AI-AAM) capable of automatically searching and designing new drug candidates from the structural formula of a known biologically active compound.

“AI-AAM” works by evaluating the binding affinity of a known active compound based on an analysis of the interactions between the compound and a set of amino acids, which are the building blocks of proteins. Furthermore, it can suggest new drug candidates with similar binding affinities to the target protein, but with different drug scaffolds*, by searching compound libraries and designing as-yet unsynthesized compounds without protein structures in place. By avoiding the costly and time-consuming process of protein structure determination, “AI-AAM” new technology can significantly speed up and increase the success rate of new drug development.

Building upon Fujifilm's extensive knowledge of molecular simulations in functional materials for such as photographic film, this newly developed “AI-AAM” is an AI technology based on a technique called “amino-acid mapping (AAM) descriptors”, which quantifies the binding energies of 20 amino acids to each compound to evaluate the binding affinity of the compound to its target protein. Using the information collected from the “AAM descriptors”, “AI-AAM” is capable of automatically searching for drug candidates from chemical libraries and also designing new drug candidates with different and stable scaffolds**. In comparison to existing AI systems which find it difficult to avoid computing synthetically unstable or unrealistic compounds, “AI-AAM” can accomplish the computation of stable compounds. Moreover, unlike common AI technology, which collects information from massive amounts of experimental data, “AI-AAM” does not require a large amount of the specific information for each disease. It is a versatile technique for drug discovery and only requires the structural formula of a known active compound to target a protein of interest. As a result, ”AI-AAM” is capable of efficient searching and designing of drug candidate compounds and is expected to increase the success rate for new drugs to enhance drug discovery and development.

[Figure1] Pictorial representations of ”AI-AAM”

Drug discovery and development including basic research, non-clinical studies, clinical studies, and filing for approval, is a long process and requires considerable amounts of time and money. Also, the probability that a drug candidate compound searched for in basic research could launch as a new drug is said to be approximately 1 in 20,000 to 30,000, and there are many cases where even compounds that bind to the target protein cannot be commercialized due to problems with toxicity. For this reason, the key to success in new drug development is to have as many compounds as possible with different scaffolds that bind to the target protein.
Currently, high-throughput screening***, which works by selecting compounds that bind to the target protein among many compounds, is commonly used, but the number of compound libraries owned by pharmaceutical companies is limited, and it is difficult to find new drug candidate compounds continuously. Recently, combining AI and methods such as docking simulation*4 that search for compounds based on the 3D structure of target proteins, as well as a process that identifies new compounds based on experimental data and the binding affinity between drug candidate compounds and target proteins, have been attracting interest for acquiring as many drug candidate compounds as possible. However, these both require analysis of the 3D structure of the target proteins, or the accumulation of data, to enhance the accuracy of AI.

Here, we report two results of “AI-AAM” chemical library searches to find new drug candidates with similar “AAM descriptors” from an anti-cancer and an antibacterial candidate compound. The newly found candidate compounds were synthesized and experimentally evaluated their biological activities, and results showed that 7% of the anti-cancer and 15% of the antibacterial compounds were actually active. This shows that the drug discovery rate of “AI-AAM” significantly exceeds that of high-throughput screening (less than 0.1%*5) and was equivalent to or slightly better than that of docking simulations (less than 10%*6).
In addition to chemical library searches, “AI-AAM” was also tested on the structural formula of an anti-cancer candidate compound in order to identify additional compound designs. Within a week, we were able to obtain a wide variety of 33 unsynthesized*7 scaffolds which supposed to have the similar binding affinities to the target protein.
Since “AI-AAM” technology is based only on the interactions between compounds and amino acids, which are smaller molecules than proteins, the computation time is expected to be less than 1/1,000 of that computing a whole target protein; thus this new technology could design multiple candidate compounds in a short period. Accordingly, “AI-AAM” is confirmed to be a highly efficient searching and designing technology in drug discovery.

Through using “AI-AAM”, Fujifilm will contribute to the rapid growth of drug development in the pharmaceutical industry and will work to create innovative drugs, both in-house and through partnering with pharmaceutical companies.

  • * The core parts of a molecular structure that are not easily changed or replaced.
  • ** From a massive number of the known compounds in chemical libraries, we have collected the information of the necessary and sufficient conditions for stable compounds. Based on the information, we developed the original AI technology that designs new compounds to fulfill the conditions. To the AI technology, we implemented the novel algorithm for solving the inverse problem, in which unknown causes are determined based on observation of their effects.
  • *** An automated experimental method for testing large numbers of chemical compounds for a specific target protein.
  • *4 A simulation method for computing the interaction between chemical compounds and the 3D structure of a specific target protein, obtained from the protein structural analysis such as X-ray crystallography.
  • *5 Source: Varma H, Lo DC, Stockwell BR. High-Throughput and High-Content Screening for Huntington's Disease Therapeutics. In: Lo DC, Hughes RE, editors. Neurobiology of Huntington's Disease: Applications to Drug Discovery. Boca Raton (FL): CRC Press/Taylor & Francis; 2011. Chapter 5. Available from: https://www.ncbi.nlm.nih.gov/books/NBK55989/
  • *6 Source: Yoshifumi Fukunishi, Seibutsu Butsuri 51 (6),252-255 (2011).
  • *7 Compounds not registered in the world's largest chemical substance database of Chemical Abstracts Service (CAS), the information division of the American Chemical Society (ACS), and “PubChem”, the chemical substance database of the U.S. National Center for Biotechnology Information (NCBI).

1. Main characteristics of “AI-AAM”

  • Taking advantage of the extensive knowledge of molecular simulations with functional materials for such as photographic films and flat panel display, “amino-acid mapping (AAM) descriptors” were developed to evaluate the binding affinity of chemical compound to its target protein by quantitatively computing the a set of binding energies of 20 amino acids to each compound. “AI- AAM” is the combination technology of simulation technique and “AAM descriptors”, for searching and designing drug candidate compounds with our newly developed AI technology.
  • By comparing “AAM descriptors” of chemical compounds and that of a known biologically active compound, ”AI-AAM” can efficiently outputs new drug candidate compounds with wide variety of scaffolds. In comparison to existing AI systems which find it difficult to avoid computing synthetically unstable or unrealistic compounds, “AI-AAM” can accomplish the computation of as-yet unsynthesized and chemically stable candidate compounds.
  • Unlike the docking simulation, one of the most major computational methods for drug discovery using 3-D structural analysis of the target proteins, and the AI technology based on enormous amount of real experimental data including binding affinities of chemical compounds to each target protein, “AI-AAM” does not depend on the complex and time-consuming experiments, and thus, it is a versatile technique for drug discovery and only requires the structural formula of a known active compound to target a protein of interest.

2. Results achieved with “AI-AAM”

(1) The compound searches of the wide range of molecular scaffolds with similar binding affinities to known biologically active compounds based on “AAM descriptors”

[Content of the Study]

  1. 1) After “AAM descriptors” were computed for 10,933 compounds*8, out of which 183 compounds have high binding affinities to the target protein, the compounds were classified into 100 groups based on the similarities of the “AAM descriptors”.
  2. 2) Using “AI-AAM”, new candidate compounds were selected for (i) an anti-cancer candidate compound from 200,000 compounds library and (ii) an antibacterial candidate compound from 100,000 compounds library.

[Results]

  1. 1) Compounds that bind to the target protein were concentrated in one group, and out of 65 compounds in the group, 34 compounds had high binding affinities to the target protein. Those compounds exhibit a wide variety of scaffolds (Figure 2).
  2. 2) In a few hours, “AI-AAM” selected 14 compounds as anti-cancer candidates from 200,000 compounds library and 13 compounds as antibacterial candidates from 100,000 compounds library. When the compounds were actually synthesized and experimentally evaluated their biological activities*9, one compound for the anti-cancer activity (drug discovery rate: 7%) and two compounds for the antibacterial activity (15%) were found.
    This shows that the drug discovery rate of “AI-AAM” significantly exceeds that of high-throughput screening (less than 0.1%) and was equivalent to or slightly better than that of docking simulations (less than 10%).
[Figure2] Structures of compounds within the same group that bonded with the target protein
  • *8 Source of the compound database: Mysinger, M. M.; Carchia, M.; Irwin, J. J.; Shoichet, B. K. J. Med. Chem. 2012, 55, 6582–6594.
  • *9 The half maximal (50%) inhibitory concentration (IC50) and minimum inhibitory concentration (MIC) were evaluated for candidate anti-cancer and antibacterial compounds.

(2) The compound designs of 33 new drug candidate compounds

[Content of the Study]
“AI-AAM” was applied to the structural formula of an anti-cancer candidate compound, and designed compounds in a week.

[Results]
“AI-AAM” could design from an anti-cancer candidate compound to obtain wide varieties of 33 unsynthesized scaffolds which supposed to have the similar binding affinities to the target protein within a week. Since “AI-AAM” technology is based only on the interactions between compounds and amino acids, which are smaller molecules than proteins, the computation time is expected to be less than 1/1,000 of that computing a whole target protein, thus this new technology could design multiple candidate compounds in a short period.

[Figure3] Pictorial representations of the “AI-AAM” designs for unsynthesized drug candidate

For inquiries on information in this media release, contact:

Media Contact:
Corporate Communications Division
TEL :+81-3-6271-2000
Other Inquiries:
R&D Division Analytical Technology Center
TEL :+81-465-74-1111